LAB-Net: LAB Color-Space oriented Lightweight Network for Shadow Removal
S | NS | ALL | ||||||
---|---|---|---|---|---|---|---|---|
RMSE | PSNR | SSIM | RMSE | PSNR | SSIM | RMSE | PSNR | SSIM |
6.65 | 37.17 | 0.9887 | 4.49 | 32.42 | 0.9727 | 4.84 | 30.49 | 0.9563 |
All the ISTD results can be found here
S | NS | ALL | ||||||
---|---|---|---|---|---|---|---|---|
RMSE | PSNR | SSIM | RMSE | PSNR | SSIM | RMSE | PSNR | SSIM |
6.56 | 35.71 | 0.9818 | 3.77 | 36.5 | 0.9813 | 4.6 | 32.22 | 0.9554 |
All the SRD results can be found here
shadow images:
![]() 105-1 |
![]() 125-1 |
![]() IMG_6581 |
![]() IMG_6916 |
results:
![]() 105-1 |
![]() 125-1 |
![]() IMG_6581 |
![]() IMG_6916 |
shadow images:
![]() 97-2 |
![]() 116-4 |
![]() IMG_6425 |
![]() IMG_6793 |
results:
![]() 97-2 |
![]() 116-4 |
![]() IMG_6425 |
![]() IMG_6793 |
shadow images:
![]() 117-15 |
![]() 124-13 |
![]() _MG_3121 |
![]() _MG_5956 |
results:
![]() 117-15 |
![]() 124-13 |
![]() _MG_3121 |
![]() _MG_5956 |
shadow images:
![]() 94-1 |
![]() 97-5 |
![]() IMG_5467 |
![]() IMG_5491 |
results:
![]() 94-1 |
![]() 97-5 |
![]() IMG_5467 |
![]() IMG_5491 |
shadow images:
![]() 99-3 |
![]() 124-4 |
![]() _MG_5771 |
![]() IMG_6456 |
results:
![]() 99-3 |
![]() 124-4 |
![]() _MG_5771 |
![]() IMG_6456 |
shadow images:
![]() 125-1 |
![]() 111-2 |
![]() _MG_5728 |
![]() IMG_6760 |
results:
![]() 125-1 |
![]() 111-2 |
![]() _MG_5728 |
![]() IMG_6760 |
We test the SBU-TimeLapse Dataset (video), USR and ADE using the model trained on ISTD.
The mask of these data is obtained by shadow detector [1].
[1] Mitigating Intensity Bias in Shadow Detection via Feature Decomposition and Reweighting
![]() shadow image |
![]() res |
![]() mask |
![]() shadow image |
![]() res |
![]() mask |
![]() shadow image |
![]() res |
![]() mask |
![]() shadow image |
![]() res |
![]() mask |
![]() shadow image |
![]() res |
![]() mask |
![]() shadow image |
![]() res |
![]() mask |
![]() shadow image |
![]() res |
![]() mask |
![]() shadow image |
![]() res |
![]() mask |
![]() shadow image |
![]() res |
![]() mask |
shadow images:
![]() 10-17 |
![]() 32-8 |
![]() _MG_2763 |
![]() IMG_6507 |
Replacing L channel of input with that of gt without changing the AB channels:
![]() 10-17 |
![]() 32-8 |
![]() _MG_2763 |
![]() IMG_6507 |
python=3.7.13
pytorch=1.12.1
pip install -r requirments.txt
- Adjust loadSize(256), FineSize(256), down_w(256), down_h(256)
- Add dataroot(ISTD trainset path), name(task name)
cd script
bash train.sh 0
0 is the gpu number
- Adjust batchs(1)
- Adjust loadSize(400), FineSize(400), down_w(128), down_h(128)
- Add dataroot(SRD trainset path), name(task name)
cd script
bash train.sh 0
0 is the gpu number
You can see the train loss:
cd script
tensorboard --logdir LAB_G_LABNet_name
- You can download our pretrained model to test.
Our ISTD checkpoint can be found here
Our SRD checkpoint can be found here
Please move the .pth to a directory to use.
- Adjust size_w(640), size_h(480), down_w(256), down_h(256)
- Add dataroot(ISTD testset path), name(task name), resroot(path to save the result)
cd script
bash test.sh 0
0 is the gpu number
- Adjust size_w(840), size_h(640), down_w(128), down_h(128)
- Add dataroot(SRD testset path), name(task name), resroot(path to save the result)
cd script
bash test.sh 0
0 is the gpu number